Individual identification of large yellow croaker (Larimichthys crocea) based on computer vision

Background Individual identification is crucial for feed nutrition and genetic breeding in fish aquaculture. Passive Integrated Transponder (PIT) tagging is currently the mainstream method for fish individual identification, but this technology has several unresolved limitations. Objective To address the invasive damage, high material costs, and low efficiency associated with PIT tagging, this study aims to develop a universal visual recognition technology applicable to fish lacking distinct phenotypic features (e.g., skin spots or stripes). Using the large yellow croaker (Larimichthys crocea), an economically important species in the East China Sea, as the validation subject, we systematically evaluated the cross-temporal-scale individual identification capability of this technology.Methods We established a fish individual identification system using a ResNet50 network architecture with residual structures as the backbone. The system learns discriminative features between individual images and constructs a recognition feature database. Results We developed an image acquisition protocol and collected a total of 7,960 bilateral images and 1,410 dorsal-ventral images from the same batch of L. crocea during their genetic breeding phase in the Baiji Bay area, filling the gap in the current individual image database for this species. The feature learning model was trained using images from 2,061 fish collected 8 weeks before spawning and tested for recognition 1 week before spawning (7 weeks later). Test results showed that the proposed method achieved a short-term recognition accuracy of 95.2% using bilateral images, with medium and long-term accuracy (across test groups) of (82.90±1.98)%. When using single-side images for medium and long-term recognition, the accuracy dropped to (77.70±3.23)% (Side 1) and 74.50±1.41% (Side 2), representing a 3–8.5% reduction compared to bilateral images, suggesting that bilateral image data should be prioritized in practical applications. Additionally, models trained on dorsal-ventral images achieved a maximum validation accuracy of only 10.78%, confirming the superior efficacy of bilateral images for biometric feature extraction and individual identification. Conclusion The individual identification technology developed in this study exhibits temporal stability unaffected by morphological changes. It provides foundational support for genetic breeding and feed nutrition research in L. crocea and offers novel insights and methodologies for individual identification in other fish species.